IBM TrueNorth : Features, Specifications. Architecture, Working, Differences & Its Applications

IBM designed small-scale prototypes from 2008 to 2010 to imitate a brain similar to neural networks. Thus, these early experiments were mostly software-based, which demonstrates the possibility of non-Von Neumann systems. After that, they launched the prototype Neuromorphic Chips in 2011, capable of simple learning & pattern recognition. After a few years of their research, they introduced a revolutionary neuromorphic chip like IBM TrueNorth. This chip has become the first digital brain processor worldwide. It operates with extreme power efficiency (70 milliwatts, signaling a breakthrough within neuromorphic engineering. This article elaborates on IBM TrueNorth, its working, and its applications.

What is IBM TrueNorth?

IBM TrueNorth is a cutting-edge neuromorphic computing chip, designed to imitate the design and working of the human brain’s neural network calculations. This is a low-power, non-von Neumann, scalable, and parallel chip that is built with 28nm CMOS technology. In addition, this neuromorphic computing chip includes neurosynaptic cores – 4096, spiking neurons – 1 million, and synapses – 256 million.

It is a significant development in computing that provides energy efficiency and high performance for robotics and image analysis tasks. Thus, its hardware architecture revolutionizes how we approach AI (artificial intelligence), machine learning, and cognitive computing by imitating the brain’s formation & function on a silicon chip.

Features

The IBM TrueNorth features include the following.

• This chip uses SNN or Spiking Neural Network, which transmits only data while a threshold is met by imitating biological neurons.
• All cores can function in parallel.
• It consumes less power compared to usual von Neumann architectures.
• Operations are event-based and asynchronous.
• Several processors can be tiled jointly for larger networks.

IBM TrueNorth Specifications

The IBM TrueNorth specifications include the following.

• IBM TrueNorth has 5.4 billion transistors
• Technology node is 28nm CMOS.
• Its power consumption in typical operation is 70 milliwatts.
• Clock speed ranges from 1 kHz to 1 MHz.
• Chip area is 430 mm²It has 4,096 neurosynaptic cores.
• It has 256 Neurons for each Core.
• Total Neurons – 1 million.
• It has 65,536 synapses for every Core.
• It has 256 million total Synapses.
• Power efficiency is~ 26 pJ/synaptic events.
• Communication is event-driven with AER (Address Event Representation)
• The programming model is Corelet programming abstraction with an ecosystem of IBM.

IBM TrueNorth Architecture

IBM Truenorth is designed with a highly parallel and event-driven architecture with different specialized components like neurons, synapses, processing elements (PEs), neural grid, synaptic memory, event-driven mechanism, programmable configuration, communication network, on-chip/off-chip memory, and power optimization.

All these components can work jointly to form this architecture to imitate biological neural networks by allowing it to process data in a highly parallel, energy-efficient, and event-driven manner. In addition, the IBM TrueNorth processor mimics the processing mechanism of the brain to provide an efficient platform for AI applications that need low-power real-time computation.

Neurons

Neurons in the IBM TrueNorth chip represent the basic units of computation. These are simple & energy-efficient digital models which are inspired by the performance of biological neurons within SNNs. Every digital model neuron is designed to imitate a biological neuron, which works in an event-driven manner.

The main function of Neurons in this chip is to generate spike signals when their inside state goes beyond a certain threshold. Thus, these signals propagate throughout the network to represent data. Each neuron in this architecture receives inputs from other neurons and generates an output spike. Thus, it updates its internal condition based on synaptic weights.

Synapses

The Synapses are the main connections between neurons, which transmit signals. Every synapse in this chip has a related weight that adjusts the power of the connection between neurons. The power of a synapse’s weight decides how much one neuron influences another neuron.

These weights can be modified eventually through learning rules like STDP (Spike-Timing Dependent Plasticity) by allowing synaptic plasticity depending on the spike timing. Thus, synapses in this chip can be digitally represented, where every synapse is equivalent to a synaptic weight. This is a numeric value that regulates the power of the connection between neurons.

Processing Elements

Processing elements or PEs are the basic building blocks in TrueNorth’s chip. Every processing element in this architecture includes a set of neurons that are in charge of executing computations in parallel. So, TrueNorth’s chip includes 4096 Processing Elements, where each element includes 256 neurons, which are organized within a 2D grid arrangement.

The main function of processing elements is to handle the neuron’s operations, like event propagation, the neuron outputs computation & spike communication. In every PE, the neurons communicate with other neurons in adjacent PEs by allowing parallel data processing.

Neural Grid

The neurons’ arrangement across the whole TrueNorth chip is known as the Neural Grid. So the neurons can be arranged in a 2D grid layout to facilitate the spike communication between neurons & across PEs.

The neural grid’s main function is to allow an efficient spatial connectivity between neurons, which allows complex computations to appear from global and local network connections. Every neuron in the neural grid can connect with thousands of others by allowing for parallel & highly solid neural networks.

Synaptic Memory

The synaptic memory in the TrueNorth chip is used to store the synaptic weights; thus, it describes how powerfully one neuron controls another neuron. Every synapse includes its memory location, which is arranged in such a way that makes it available for proficient read/write operations. This memory can be arranged hierarchically across various chip levels with local memories connected to processing elements & global memory shared across the whole system.

Event-Driven Mechanism

The TrueNorth chip has an event-driven mechanism where only neurons fire when they get enough input to activate a signal. So this is a significant difference from usual computing architectures that function in a clock-driven manner. This event-driven mechanism contributes to the energy efficiency of the system because neurons are active only when they are involved in processing, reducing needless energy consumption.

Programmable Configuration

TrueNorth chip allows for the arrangement of the synapses, neurons & the learning rules that rule them. Thus, this flexibility allows the system to be adapted to specific applications or tasks like sensory processing, autonomous decision-making, or even pattern recognition. This chip supports learning mechanisms like synaptic plasticity rules and STDP that adjust synaptic weights depending on the patterns of signals within the network.

Communication Network

The communication network helps the transmission of the spikes between neurons across different processing elements. Thus, every processing element is independent to process neurons in parallel, and a strong communication network ensures that spikes are efficiently transmitted across the processor. This communication network can be designed to support high-speed communication & low-latency signal transmission by ensuring that computations take place in real time.

On-Chip and Off-Chip Memory

The on-chip memory of this chip stores the synapses, neurons & the network configurations. Thus, this memory is strongly integrated into the architecture for low-latency and efficient access operation. The off-chip memory can also be available for larger-scale applications to store wider models, data training & extra network parameters that can be used throughout computation.

Power & Performance Optimization

The TrueNorth processor design emphasizes very-low power consumption, making it mostly suitable for mobile devices or embedded systems that need power efficiency. This processor functions in the milliwatt range related to the brain’s power consumption. Neurons consume power while they are energetically firing, thus significantly decreasing the required energy for computation. This is mainly significant for ML and AI tasks that need real-time and continuous processing.

Chip-Level Architecture

Each TrueNorth processor includes 256 million synapses and one million neurons with 4096 processing elements. It is arranged in such a way that permits the system to increase simply, including more chips as required for more complex and larger neural networks. It is achievable to produce a massive and distributed neural network that handles complex tasks & processes large amounts of data by connecting various TrueNorth chips.

TrueNorth Chip Software Ecosystem

The TrueNorth chip can be programmed and worked with a specialized software ecosystem developed by IBM known as the Corelet Programming Environment. The key components of this software system are discussed below.

Corelets

Modular building blocks like corelets are similar to NN components. Every Corelet summarizes the program of a neural network that can be arranged on the TrueNorth chip. These building blocks are written in a DSL (domain-specific language) based on Python.

Compass

Compass is the SDK (software development kit) for TrueNorth. It provides tools for writing, simulating, debugging & compiling code for operation. It is written in C++ and Python, which provides a simulation on fixed hardware before it is uploaded to TrueNorth hardware.

IBM Eedn

IBM Eedn is an energy-efficient deep neuromorphic network, a high-level framework that changes into TrueNorth-compatible models from deep learning networks. It changes convolutional neural networks (CNNs) into spiking NNs.

How does IBM TruNorth Work?

The TrueNorth neuromorphic processor works by imitating the brain’s neural design & function like the cortex. This can be achieved by an event-driven and massively parallel architecture through programmable neurons & synapses. In addition, the TrueNorth processor doesn’t depend on a central clock or sequential instructions processing, like traditional computers, but instead on the flow between consistent neurons.

Generally, the core architecture of IBM Truenorth mainly includes Neurosynaptic Cores, Neurons, Synapses, and Network-on-Chip. The working of this processor involves different steps, like the following.

First, the data in event-driven Processing can be processed through the spike flow between neurons by mimicking the way neurons converse within the brain. The processor in parallel processing allows massive parallelism using neurons and cores, which operate simultaneously. The nonexistence of a global clock permits asynchronous operation, which leads to greater flexibility and lower power consumption. Finally, the scalability of this processor allows the combination of various chips to manage more complex and larger tasks.

IBM Truenorth vs Intel Loihi

The neuromorphic chips like IBM’s TrueNorth & Intel’s Loihi are designed to imitate the human brain to execute computations. So the difference between IBM Truenorth vs Intel Loihi includes the following.

Advantages

The advantages of IBM TruNorth include the following.

• IBM TrueNorth chip can be optimized for event-driven and low-power processing to use in edge computing applications like sensor data analysis and IoT devices.
• It consumes less power; thus, it is suitable in energy-constrained environments and battery-powered devices.
• The processor can be scaled up by merging various chips to manage more complex tasks.
• Its architecture tolerates failures and faults within individual components.
• It doesn’t need external memory access.
• This chip decreases latency and enhances energy efficiency significantly.
• It is compact and lightweight.
• In addition, its neurosynaptic cores with neurons & synapses allow high-speed & parallel computation, which is related to the data processing in a brain.

Disadvantages

The disadvantages of IBM TruNorth include the following.

• IBM’s TrueNorth uses nanometer manufacturing processes; thus, analog circuitry implementation is very difficult at these small sizes.
• It occupies more space.
• The memory density in the Von-Neumann architecture is higher.
• This chip functions at much lower accuracy by limiting its applicability to precision-sensitive tasks.
• The real trouble with TrueNorth is that these chips are esoteric from frameworks and programming positions, thus clicking numerous together has been a challenge from a development standpoint.

IBM TrueNorth Applications

The applications of IBM TruNorth include the following.

• TrueNorth processor analyzes video feeds to recognize objects like cars, bicycles, and pedestrians in real-time by using very little power.
• This chip can be used for natural language processing and speech recognition wherever understanding patterns is crucial in audio & text.
• This chip analyzes medical images from EEG tests and fMRI scans, which potentially helps in diagnosis & research.
• In addition, this processor shines at performing inference, particularly at the edge, while not being perfect for training large deep learning models.
• It identifies patterns within large datasets and is helpful in fraud detection & cybersecurity areas.
• It is suitable for improving the decision-making and learning abilities of robots due to low power consumption & capability to process sensory data.
• This chip is used in powering edge devices like wearables and smartphones due to its low power consumption.
• TrueNorth chip can be used in drones and remote sensors for processing & analysis of data at the edge. In addition, it reduces the need for stable communication through a central server.
• It monitors shipping lanes for potential hazards and anomalies.
• In addition, this chip can be used in the future with sensors within medical devices to recognize bacteria.

Thus, this is an overview of a neuromorphic computing chip like IBM TrueNorth, where its digital neuromorphic architecture is manufactured with a typical semiconductor process. TrueNorth chip signifies a breakthrough system that powerfully harnesses abilities inspired through neural systems. In addition, is a completely new creation of a very-low power parallel architecture which is unmatched within computational efficiency for mainly stochastic sensory-cognitive workloads. Here is a question for you: What is Brainchip Akida?